AI continues to dominate the business conversation worldwide, with organisations heavily investing in industry-leading AI talent, data strategies and algorithms. However, it’s clear that more than aspiration is required to deliver on AI initiatives.

Turning innovation into tangible action has been difficult for some. According to S&P Global Market Intelligence, the share of businesses scrapping most of their AI initiatives rose to 42% – up from 17% last year.

The problem, said VIRTUS Data Centres chief revenue officer Darren Watkins, is the lack of physical environments capable of sustaining AI workloads.

“Many organisations are treating AI as a software or data problem, when in reality it is an infrastructure challenge,” he told Capacity. “The biggest oversight we see is the assumption that existing digital environments can simply absorb AI workloads.”

The problem with legacy data centre infrastructure

Running AI on a large scale requires purpose-built data centres that have the specific, highly advanced infrastructure required for the technology’s density and cooling.

Watkins explains the impact of this: “Training a large model or running inference continuously at scale demands enormous, sustained compute power. It places unprecedented stress on power delivery, cooling systems and the physical fabric of a facility.

“Companies often underestimate just how far AI’s hardware and thermal profiles exceed traditional IT.”

Legacy enterprise and colocation facilities have been built for stability and predictability. However, AI training can exceed fifty or more kilowatts per rack, Watkins said, meaning that existing systems often cannot handle the demand.

“Cooling systems designed for air alone cannot remove the heat generated by GPU clusters. Power distribution paths become bottlenecks, and the weight of liquid cooled racks often exceeds structural limits. Even cable management and airflow containment must be re-engineered,” he added.

“In short, these workloads need a completely different architecture – one that is optimised for density, precision cooling and resilience. Without it, performance, efficiency and uptime all suffer.”

Another blind spot businesses can have is sustainability, Watkins explained. As AI consumes significant amounts of energy, efficiency and responsible sourcing is critical.

“Businesses that fail to integrate sustainability into their AI strategy risk falling short of investor expectations and upcoming regulation,” Watkins said.

A ‘new type’ of data centre

Running AI requires a data centre that is engineered specifically for high-density compute. This means it contains liquid cooling, high-capacity power distribution systems, reinforced structures to support additional load and embedded sustainability into the design.

“These features together allow AI workloads to run continuously and efficiently, without compromising the environment or operational resilience,” Watkins explained.

He added that many enterprises have already invested heavily in data science teams, only to discover their physical capacity can’t keep pace.

“We’re seeing a shift from curiosity to urgency,” he said. “The conversations that resonate most are around readiness and risk. Businesses want clarity on what it takes to move from pilot to production. This means looking at how to secure the right power, cooling and connectivity without overcommitting capital.

“They also want partners who understand both the technical and the operational sides of the equation – the ability to deliver high-density environments that are sustainable, scalable and compliant.”

What VIRTUS sees as key enterprise challenges

Watkins noted that VIRTUS sees many clients thinking the challenge is compute access, but said it’s actually about balancing compute with the right supporting infrastructure.

“You can have the most advanced GPUs in the world, but if your power and cooling systems can’t keep them running at full capacity, performance and ROI both collapse,” he added.

He notes another misconception around location: “AI performance is directly linked to proximity to data sources, networks and users. As latency becomes a commercial issue, the strategic placement of data centres is as critical as their technical design.”

According to VIRTUS, flexibility is also being underestimated, given the speed of AI architectures evolving, as Watkins explained: “The ability to reconfigure density, cooling, or layout without downtime can make the difference between scaling successfully and starting over.”

With this in mind, VIRTUS views modularity as the key to balancing capital investment in liquid cooling, power distribution and thermal management against the risk that AI architectures might evolve in unexpected ways.

“We’re investing in flexible design frameworks that allow us to adapt to changing density and cooling requirements over time,” Watkins said.

According to Watkins, the frameworks include:

• Modular cooling loops that can support different technologies.
• Power architectures designed to scale incrementally as demand grows.
• Hall configurations that allow rapid re-fit without affecting other customers.

The company is also taking a long-term view on sustainability and efficiency, particularly as AI evolves.

“The underlying principles of efficient design (short power paths, advanced monitoring, circular heat use, renewable sourcing) remain relevant whatever form AI infrastructure takes,” Watkins said.

“The foundations of resilient, efficient, high-capacity design are enduring. Our role as an operator is to make sure that our customers can innovate freely within that framework.”

Related stories

techoraco announces the return of Datacloud Energy Europe 2026

AMD, Cisco, HUMAIN: Partnering to advance AI infrastructure

Nvidia’s Blackwell dilemma explained: China, chips and the future of AI infrastructure